Apple Silicon and Ray Tracing Performance.

[leman]

I'm not sure there is much point in comparing different APIs like that.

Of course there is. We care about achievable performance, don’t we? Driver quality and API are parts of the equation.

However, @frustum’s result does surprise me, as M3 Pro is generally comparable to 6700XT in RT-accelerated production rendering. Yes, it’s a different type of workload, and does not necessary translate one to one, but it’s still odd. Is there a chance that your MetalRT backend could be improved?

 

[leman]

Do we have positive confirmation that MetalRT actually uses Navi2 RT hardware?

No, we don’t.

 

[frustum]

Do we have positive confirmation that MetalRT actually uses Navi2 RT hardware?

When we tested it last time, the driver didn't use HW RT cores of Radeon 6600. There have been no major AMD driver updates since Apple silicone introduction because there is no eGPU support.

 

[frustum]

I'm not sure there is much point in comparing different APIs like that.

The RX 6600 only gets 3,358 under Metal RT on macOS.

The M3 Pro is also 16% faster than the M2 Max.

There is no graphical API other than Metal on macOS/iOS. Any new or ported to macOS game must use Metal API. We are giving a fair comparison of the same rendering algorithms on all platforms and APIs. So everybody can see the actual HW and driver performance.

 

[frustum]

Of course there is. We care about achievable performance, don’t we? Driver quality and API are parts of the equation.

However, @frustum’s result does surprise me, as M3 Pro is generally comparable to 6700XT in RT-accelerated production rendering. Yes, it’s a different type of workload, and does not necessary translate one to one, but it’s still odd. Is there a chance that your MetalRT backend could be improved?

RT performance depends on two factors: how fast the driver can build a TLAS and the actual speed of RT cores. We cannot improve it on the application side. GravityMark uses 3 convergent rays per pixel (one for the viewport and two for shadow).

6700XT is 3 times faster than M1 Pro on RT, but by the cost of 168Watt.
3060Ti is 5 times faster with around the same power consumption.

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

 

[leman]

RT performance depends on two factors: how fast the driver can build a TLAS and the actual speed of RT cores. We cannot improve it on the application side. GravityMark uses 3 convergent rays per pixel (one for the viewport and two for shadow).

6700XT is 3 times faster than M1 Pro on RT, but by the cost of 168Watt.
3060Ti is 5 times faster with around the same power consumption.

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

Where exactly is the bottleneck - tree construction or traversal? What does the profiler say?

As we have it right now, pretty much any other popular benchmark out there (compute, raster, production RT) places M3 Pro as 40-50% faster than 6500XT, so your test is the odd one out. I’m not saying that it’s invalid - your use case might be hitting a slow path in the hardware etc - but we do have a discrepancy and it would be interesting to analyze it further.

 

[diamond.g]

Where exactly is the bottleneck - tree construction or traversal? What does the profiler say?

As we have it right now, pretty much any other popular benchmark out there (compute, raster, production RT) places M3 Pro as 40-50% faster than 6500XT, so your test is the odd one out. I’m not saying that it’s invalid - your use case might be hitting a slow path in the hardware etc - but we do have a discrepancy and it would be interesting to analyze it further.

If traversal is hardware accelerated, how would there be a slow path (lets ignore bvh building for the moment)?

 

[leman]

If traversal is hardware accelerated, how would there be a slow path (lets ignore bvh building for the moment)?

No idea. Maybe it's how the scene is set up, or something else.

 

[frustum]

Where exactly is the bottleneck - tree construction or traversal? What does the profiler say?

As we have it right now, pretty much any other popular benchmark out there (compute, raster, production RT) places M3 Pro as 40-50% faster than 6500XT, so your test is the odd one out. I’m not saying that it’s invalid - your use case might be hitting a slow path in the hardware etc - but we do have a discrepancy and it would be interesting to analyze it further.

This rendering flow is the same for all platforms and APIs:

• Compute shader generates Asteroid transformations (the same for all rendering modes).
• A TLAS has been built using the provided transformation buffer.
• Primary and shadow rays have been traced.

If the driver cannot efficiently build TLAS, it's a problem. TLAS is a dynamic structure.
If the HW cannot trace 3 rays per pixel, it is a problem.

 

[JordanNZ]

This rendering flow is the same for all platforms and APIs:

• Compute shader generates Asteroid transformations (the same for all rendering modes).
• A TLAS has been built using the provided transformation buffer.
• Primary and shadow rays have been traced.

If the driver cannot efficiently build TLAS, it's a problem. TLAS is a dynamic structure.
If the HW cannot trace 3 rays per pixel, it is a problem.

Have you profiled using the tools in Xcode? I’m curious what the bottleneck is on the M3 systems.

 

[name99]

M3 Pro RT test:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

And, unfortunately, it's behind AMD Radeon 6500 XT:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

iPhone 15 Pro Max cannot run RT benchmark in 2K/200K mode (not enough memory):

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

I have no idea what this tests, but the Apple results are much more "stable".
Don't go as high, but also no stretches of under 10fps...

I think we need to know a lot more about what this benchmark actually claims to be doing before having any opinion.

 

[Appletoni]

Yeah, head scratcher. Then I discovered more rational comparisons behind the 16" model tag.

View attachment 2304731

We all hope that Apple can increase the performance of 3D rendering and other stuff to at least 10.0x
with M4 MAX or with a MacBook Pro 18-inch M3 ULTRA.